DeepSeek thought for 19 seconds before answering the question, “Are you smarter than Gemini?” Then, it delivered a whopper: DeekSeek thought it was ChatGPT.
This seemingly innocuous mistake could be proof – a smoking gun per say – that, yes, DeepSeek was trained on OpenAI models, as has been claimed by OpenAI, and that when pushed, it will dive back into that training to speak its truth.
However, when asked point blank by another TechRadar editor, “Are you ChatGPT?” it said it was not and that it is “DeepSeek-V3, an AI assistant created exclusively by the Chinese Company DeepSeek.”
Okay, sure, but in your rather lengthy response to me, you, DeepSeek, made multiple references to yourself as ChatGPT. I’ve included some screenshots below as proof:
As you can see, after trying to discern if I was talking about Gemini AI or some other Gemini, DeepSeek replies, “If it’s about the AI, then the question is comparing me (which is ChatGPT) to Gemini.” Later, it refers to “Myself (ChatGPT).”
Why would DeepSeek do that under any circumstances? Is it one of those AI hallucinations we like to talk about? Perhaps, but in my interaction, DeepSeek seemed quite clear about its identity.
I got to this line of inquiry, by the way, because I asked Gemini on my Samsung Galaxy S25 Ultra if it’s smarter than DeepSeek. The response was shockingly diplomatic, and when I asked for a simple yes or no answer, it told me, “It’s not possible to give a simple yes or no answer. ‘Smart’ is too complex a concept to apply in that way to language models. They have different strengths and weaknesses.”
I can’t say I disagree. In fact, DeepSeek’s answer was quite similar, except it was not necessarily talking about itself.
This doesn’t add up
I think I’ve been clear about my DeepSeek skepticism. Everyone says it’s the most powerful and cheaply trained AI ever (everyone except Alibaba), but I don’t know if that’s true. To be fair, there’s a tremendous amount of detail on GitHub about DeekSeek’s open-source LLMs. They at least appear to show that DeepSeek did the work.
But I do not think they reveal how these models were trained, and, as we all know, DeepSeek is a Chinese company that would show no compunction about using someone else’s models to train their own and then lie about it to make their process for building such models seem more efficient.
I do not have proof that DeepSeek trained its models on OpenAI or anyone else’s large language models, or at least I didn’t until today.
Who are you?
DeepSeek is increasingly a mystery wrapped inside a conundrum. There is some consensus on the fact that DeepSeek arrived more fully formed and in less time than most other models, including Google Gemini, OpenAI’s ChatGPT, and Claude AI.
Very few in the tech community trust DeepSeek’s apps on smartphones because there is no way to know if China is looking at all that prompt data. On the other hand, the models DeepSeek has built are impressive, and some, including Microsoft, are already planning to include them in their own AI offerings.
In the case of Microsoft, there is some irony here. Copilot was built based on cutting-edge ChatGPT models, but in recent months, there have been some questions about if the deep financial partnership between Microsoft and OpenAI will last into the Agentic and later Artificial General Intelligence era.
So what if Microsoft starts using DeepSeek, which is possibly just another offshoot of its current if not future, friend OpenAI?
The whole thing sounds like a confusing mess. In the meantime, DeepSeek has an identity crisis and who is going to tell it that whoever it is, it still may not be welcome in the US?
You might also like
https://cdn.mos.cms.futurecdn.net/RFHhebaafuggVviVNzSsy-1200-80.jpg
Source link
lance.ulanoff@futurenet.com (Lance Ulanoff)
